File size: 2,493 Bytes
a40cb3b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
import torch
import gradio as gr
import networkx as nx
import matplotlib.pyplot as plt
from transformers import GPT2Model, GPT2Tokenizer
from sklearn.cluster import KMeans

# 1. Load a real small model
device = "cuda" if torch.cuda.is_available() else "cpu"
model_name = "gpt2" # 124M parameters
tokenizer = GPT2Tokenizer.from_pretrained(model_name)
model = GPT2Model.from_pretrained(model_name).to(device)

def get_hidden_state(sequence_str):
    inputs = tokenizer(sequence_str, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model(**inputs, output_hidden_states=True)
    # Use the last hidden state of the last token
    return outputs.hidden_states[-1][0, -1, :].cpu().numpy()

def analyze_dfa(input_text):
    """
    Simulates a 'State Probe'. 
    Input: 'Right, Up, Left'
    Logic: Generates a graph showing how the model's internal representation 
    changes with each move.
    """
    moves = [m.strip() for m in input_text.split(",")]
    history = ""
    states_vectors = []
    
    # Track the "path" through the model's internal space
    for move in moves:
        history += f" Move {move}."
        vec = get_hidden_state(history)
        states_vectors.append(vec)
    
    # Clustering: Vafa's Compression metric
    # We cluster activations to see which moves the model thinks are 'equivalent'
    num_clusters = min(len(moves), 4)
    kmeans = KMeans(n_clusters=num_clusters, n_init=10).fit(states_vectors)
    labels = kmeans.labels_
    
    # Build the DFA Graph
    G = nx.DiGraph()
    for i in range(len(moves)-1):
        u, v = f"S{labels[i]}", f"S{labels[i+1]}"
        G.add_edge(u, v, label=moves[i+1])
    
    # Draw the DFA
    plt.figure(figsize=(6, 4))
    pos = nx.spring_layout(G)
    nx.draw(G, pos, with_labels=True, node_color='lightblue', node_size=2000)
    edge_labels = nx.get_edge_attributes(G, 'label')
    nx.draw_networkx_edge_labels(G, pos, edge_labels=edge_labels)
    
    plt.savefig("dfa_plot.png")
    return "dfa_plot.png", f"Found {num_clusters} distinct internal states."

# 3. Gradio Interface
demo = gr.Interface(
    fn=analyze_dfa,
    inputs=gr.Textbox(placeholder="Enter moves separated by commas, e.g.: Right, Up, Left, Down"),
    outputs=[gr.Image(label="Extracted Model DFA"), gr.Text(label="Analysis")],
    title="World Model DFA Extractor",
    description="This tool probes GPT-2's internal activations to see if it treats different move sequences as the same 'State'."
)

demo.launch()